This article belongs to the super dry goods methodology. Whether it is a product, operation or data analysis practitioner, as long as it is a content-based product form, it will need to be exposed to the feed stream, and the entire work almost revolves around the theme of content optimization. In this article, I will tell you a summary of your understanding of this methodology. I believe that reading this article will help you.
1. What is the feed flow?
Feed流是将若干消息源组合在一起，帮助用户持续地获取最新的内容。我们无需主动搜索，自动呈现琳琅满目的内容。它对我们了如指掌，给我们想了解的，让我们不停的刷新。我们熟知的微博、知乎、今日头条、微信朋友圈、各类短视频等都是feed流的展示模式。我们以今日头条为例，Feed流如图：
After reading the picture, I believe that everyone has a basic understanding of the feed stream. Its core is personalized recommendation, that is, through various strategies, some content is filtered from the content pool, and then displayed to users after the strategy is sorted.
Second, how to generate feed stream?
The generation of the feed stream follows the four steps of strategy formulation: question-> input-> calculation-> output, that is, in order to show the user the appropriate content, a series of data indicators are input, logical calculation is performed, and finally a user satisfaction is output. Feed stream results.
** 1, ** The recommended input dimension for content
We can consider input indicators from the three dimensions of content, users, and environment;
Content : Today's Toutiao is a comprehensive information aggregation platform, covering graphics, videos, small videos, micro-headlines, Q & A, etc. Each content has its own characteristics and vertical categories, you need to consider how to extract the characteristics of different content types to do well recommend.
** User: ** Cover user basic information, occupation, age, gender, etc., as well as interest preference tags based on the user's past behavior data, and even get the user's device information and behavior information on other apps.
** Environment: ** Users will have specific preferences for information in different scenarios such as work, home rest, commuting, travel, etc., depending on the surrounding environment.
2. User characteristics
3. Commonly used matching data features
4. Special artificial strategies
With the above input features and logic calculation and generation strategy, you can output a feed stream result. So how to judge whether the user is satisfied with this feed stream?
3. Evaluation of the effect of feed flow
A basic principle is that if you want to evaluate the performance of the feed stream, you need to score it through various aspects, so as to find out how much the feed stream is liked by users. Scoring can be considered from the two dimensions of ranking and content itself, that is, the more content the user likes and the more content the user is interested in, the better the feed flow effect. Refinement evaluation indicators can consider the following data dimensions:
1. Traffic: Top N traffic
2. Click-through rate (front N-brush click-through rate, overall click-through rate): the most intuitive data, the higher the click-through rate of the feed stream clicked by the user, the more interesting the recommended content is. CTR says that the user ’s favorite content is higher
3. Duration of stay: The longer the user stays in the content of the feed stream, the more the content of the feed attracts the user (excluding the interference of the title party caused by the false high ctr)
4. Reading Completion: Similar to the concept of length of stay, high completion indicates that the content meets the user
5. Activity: users like, comment, share, follow, collect and other behaviors
With the evaluation indicators, you can set the weight scores for each indicator segment to calculate the total score of a feed stream. The following behavior weight score example:
[External chain image transfer failed, the source site may have an anti-theft chain mechanism, it is recommended to save the image and upload it directly (img-KlxShHnt-1587523052985) (https://upload-images.jianshu.io/upload_images/8429180-2a6e6bc35bb0c222?imageMogr2 / auto-orient / strip% 7CimageView2 / 2 / w / 1240)]
Fourth, the optimization strategy of the feed stream
Through the above steps, we have initially generated a feed stream and its score, the following need to continue to optimize iteration. Let's take today's headline as an example to see what is the biggest problem with its feed stream? Based on these problems, let's see what solutions can solve these problems and optimize the feed flow. So we use sampling analysis method to sample the articles in the recommended feeds of different users, conduct survey and evaluation, analyze all kinds of problems badcase and find solutions. The following is the result of sampling analysis:
** Homogeneity of content: ** From the perspective of content production, the content released by the platform creators is becoming more and more similar. From the perspective of readers, the content pushed every day is not fresh enough, they are all of the same type. Toutiao is an information platform. Homogeneous content will greatly reduce the user experience and affect the core competitiveness of Toutiao, and this part of the problem accounts for nearly 40%. It is urgent to optimize and has a higher priority.
** Low content quality: ** Some content will hang the names of various gangsters as headlines, but when you click in, you will find that the content of the article is of very low quality. Different users also perceive content differently. If users feel that the quality of the content they see is low, and they click on several articles in a row to see most of the low-quality content, users will choose to jump out of the platform.
** Old news: ** The labels of different channels and different levels must be set with fine-tuned time limits. If users can still see the unknown old news in the main feed a few months ago, they may think that the content of the platform is not enough. Unable to get first-hand information in time.
** Missing content tag: ** The tag definition is too few to focus. If it is a high-frequency word tag, it will result in matching a large amount of irrelevant content, and if a low-frequency word tag, it may not match enough content.
** Duplicate topics: ** Hot topic articles are indispensable, but if the duplicate topics are too heavy, users will feel that the platform is too monotonous and need to control the frequency of articles with the same topic.
Through the analysis of various problems, we can formulate a preliminary optimization plan according to the corresponding problems:
推荐内容优化是个持续的过程，需要综合各个维度、多个角色，不断的优化迭代，评估再优化迭。且产品每个阶段，问题的类型、解决问题的方案都会不同。此外，在制定各类标准时，也要将产品的调性纳入考量因素，这样才是真正的用户导向。
The above is my understanding of the feed stream generation, performance evaluation and optimization methodology of content-based products. Welcome to pay attention to my WeChat public account, and exchange data analysis issues at any time.
Service recommendation
